Case Study

Using a Development-Set Check to Separate Search Problems from Metric Problems

Case context: A translation model is examined on one development example. For the evaluation score Eval()Eval(\cdot), the human reference translation TT^* receives a value of 4, while the model’s prediction TpredT_{pred} receives a value of 9.

Question: Using the Optimization Verification check, how should you label this error, and what does the score comparison imply about the model’s search process and the evaluation metric?

Sample answer: This should be labeled a metric problem because Eval(T)Eval(Tpred)Eval(T^*) \le Eval(T_{pred}) ($4 \le 9). That means the model’s search procedure found an output that the metric prefers over the reference translation, so the metric is failing to favor the correct answer rather than the search procedure failing to optimize it.

Key points:

  • Label the error as a metric problem.
  • Note that the conditionEval(T^*) \le Eval(T_{pred})is satisfied since $4 le 9.
  • Conclude that the evaluation metric is the source of the issue, not the model’s search process.

Rubric: The response must identify the error as a metric problem, cite the relevant inequality condition (Eval(T)Eval(Tpred)Eval(T^*) \le Eval(T_{pred})), and explain that the metric is at fault rather than the optimization or search procedure.

0

1

Updated 2026-08-12

Contributors are:

Who are from:

Tags

Machine Learning

Deep Learning

Supervised Learning

Dive into Deep Learning @ D2L

Data Science

Machine Learning Strategy

Machine Learning Yearning @ DeepLearning.AI

Related